technical report meteoswiss no. 255 · contact: [email protected] method information...
TRANSCRIPT
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:
User manual and documentationSophie Fukutome and Anne Schindler
Recommended citation:
Fukutome S, Schindler A : 2015, MeteoSwiss extreme value analyses:
User manual and documentation, Technical Report MeteoSwiss, 255, 50 pp.
Editor:
Federal Office of Meteorology and Climatology, MeteoSwiss, c© 2015
MeteoSwiss
Operation Center 1
8058 Zurich-Flughafen
T +41 58 460 91 11
www.meteoswiss.ch
ISSN: 2296-0058
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation
Sophie Fukutome and Anne Schindler
MeteoSwiss extreme value analyses:User manual and documentation 5
Abstract
Extreme value analyses of meteorological quantities, such as precipitation, or temperature, are of
great importance in design engineering and risk management. At the end of 2013, MeteoSwiss has
made extreme value analyses of precipitation publicly available for 27 stations of its observational
network. The present report provides users with a manual for understanding and interpreting these
analyses; it documents the data used and the statistical methods employed. It is meant for engineers
and other experts familiar with basic statistical concepts. The report is structured in three parts. The
first describes and explains the contents of the analysis sheets, the second documents the underlying
data, and the third is dedicated to the theoretical background.
Zusammenfassung
Extremwertanalysen von meteorologischen Grossen, wie Niederschlag oder Temperatur, sind von
grossem Interesse fur Dimensionierung und Risikomanagement. MeteoSchweiz hat Ende 2013 Ex-
tremwertanalysen von Niederschlag, gemessen an 27 Stationen des meteorologischen Netzwerks der
MeteoSchweiz, der Offentlichkeit zur Verfugung gestellt. Dieser Bericht liefert eine “Bedienungsan-
leitung”, um diese Analysen zu verstehen und zu intepretieren und dokumentiert die zugrundeliegen-
den Daten und Methoden. Er richtet sich an Ingenieure und andere Experten, die statistisches Grund-
lagenwissen besitzen. Der Bericht ist dreiteilig gegliedert: Das erste Kapitel beschreibt und erklart die
Inhalte der Analyse-Blatter. Das zweite Kapitel dokumentiert die zugrundeliegenden Daten und das
dritte Kapitel befasst sich mit den notigen theoretischen Grundlagen.
Technical Report MeteoSwiss No. 255
6
ContentsAbstract 5
1 Introduction 7
2 Interpretation 10
2.1 Return level plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Reliability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Example stations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.1 Geneve-Cointrin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.2 Zurich / Fluntern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3 Data 19
3.1 Stations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 Data quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 Heavy precipitation data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4 Theoretical background: Extreme Value Statistics 24
4.1 The GEV distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.2 Inference for the GEV distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2.1 Maximum likelihood estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.2.2 L-moments estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3 Assessing the fit of the model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.3.1 Visual guides . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.3.2 Goodness-of-fit criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.4 Return levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.4.1 Return level estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.4.2 Uncertainty estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.5 Extreme Value Analysis example . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Abbreviations 41
List of Figures 43
List of Tables 46
References 47
Acknowledgment 48
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation1 Introduction
7
1 Introduction
Switzerland has witnessed a number of weather events causing devastating floods in the last decades,
as for instance in August 2005, when a large part of the Swiss northern Alpine rim was affected.
Such events raise questions regarding adequate planning of infrastructure based on reliable statistical
estimates. The most recent official documents on the matter were compiled by the Federal Institute for
Forest, Snow and Landscape Research (WSL) in the 70s and 80s, and are outdated, both in terms of
precipitation data and of statistical methods.
As a federal office operating a broad network of meteorological observation stations, MeteoSwiss has
a public responsibility to provide information on the severity of rare events. In a first step, a web page
(http://www.meteoswiss.ch/home/klima/vergangenheit/extrem-niederschlaege.html) was
set up on the website of MeteoSwiss at the end of 2013, presenting extreme value analyses of 1-day, 2-
day, and 3-day precipitation at 27 stations of the MeteoSwiss observational network. On the instigation
and with the support of the Federal Office for the Environment (FOEN), a more comprehensive platform
is underway with analyses at more than 350 (60) stations for daily (subdaily) precipitation, in which
state-of-the-art statistical methods will be applied.
The extreme value analyses made available in 2013 consist of a 2-page sheet in pdf-format. They
are intended for engineers and other experts requiring quantitative information for design and risk
management. The present report explains how to read and interpret these analyses, and documents
the underlying data and methods. The methods employed here are long established, and climatological
aspects are not addressed. Thus, the report must be seen as a ”user manual” for a person familiar with
basic statistical concepts, wishing to understand the contents of the analyses, and how they came into
being. The exact mathematical expressions have been added for completeness in separate boxes, but
can be ignored, as they are not essential for understanding the text.
The report is divided into three parts. The first provides a detailed description (chapter 2) of the con-
tents of the analysis sheet. The second is dedicated to the data used (chapter 3), their measurement,
eventual manipulation, and quality issues. Finally, the third part (chapter 4) describes the theoretical
background: the relevant aspects of extreme value statistics, the assumptions that need to be made,
and when these may be considered appropriate in this particular context.
This report will be updated as more sophisticated methods are introduced for the new web platform.
”Graphical” table of contents (pages 8 - 9):
Example of a 2-page extreme value analysis sheet (as appears on the web page).
Technical Report MeteoSwiss No. 255
8
Zurich / Fluntern: 556m, 47.38N, 8.57E
Extreme Value Analysis1-day precipitation, 5:40-5:40 UTC
1961 - 2013 (number of missing years: 0)
Block Maxima (GEV). Reliability of results: questionable.
Plot of return levels and their uncer-tainty (ordinate) for a given returnperiod (abscissa).
The best estimate of the return levelsis colored blue. The return level 95%confidence intervals are in green.
The dots represent observations (inthe ordinate) to which empirical return pe-riods (abscissa) have been assigned. Theseempirical return periods depend solely onthe sample size. Here, for instance, thelargest observation is automatically assigneda return period of approx. 53 years.
return period [years]
retu
rn le
vel [
mm
/day
]
1.1 2.33 5 10 20 50 100 200
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●5010
015
020
0
© MeteoSwiss
Table of the largest annual extrema inthe period analysed.
If two large events occur in the sameyear, only the largest will appear in thistable. The return periods are estimatedwith the fitted GEV distribution.
date precipitation[mm/day]
estimated returnperiod [years]
1968-09-21 103.1 342007-08-08 97.8 271978-08-07 97.2 271973-06-23 97.1 271999-05-12 94.2 24
Table of return levels for a selection ofreturn periods.
The 95% confidence intervals are shown inbrackets.
return return confidenceperiod value interval[years] [mm/day] [mm/day]
2 48.2 ( 44.6 - 52.6 )5 63.2 ( 56.3 - 71.5 )
10 75.8 ( 64.5 - 90.9 )20 90.4 ( 72.3 - 117.6 )30 100.0 ( 76.2 - 136.8 )50 113.6 ( 81.4 - 169.5 )
100 135.1 ( 88.7 - 233.5 )
©MeteoSwissKrahbuhlstrasse 58, 8044, Zurichrre150yx; xs 2.3.1; June 10, 2015
Contact: [email protected]
Short summary
Interpretation ..... p. 14
Return level plot
Interpretation ..... p. 10
Definition ........... p. 37
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentationgraphical table of contents
9
Distribution Function and estimation methods
• The Generalized Extreme Value distribution (GEV) is fitted to the yearly extrema.
• The distribution parameters are estimated with Maximum Likelihood.
• The confidence intervals are estimated with parametric resampling.
Data and data quality
• The raw data is quality-checked, but not homogenized.
• Missing data: None.
Distribution Parameters
• Location: 44.25
• Scale: 10.39
• Shape: 0.25
Additional information
Time series of extrema
time
mm
/day
1960 1970 1980 1990 2000 2010
3040
5060
7080
9010
0
Histogram of Extrema
mm/day
dens
ity
40 60 80 100
0.00
0.02
0.04
0.06
0.08
● ●●●●●●●
●●●●
●●●●●●
●●●●●
●●●●●●
●●●●●●●●●●●
●●●
●
●
●
●
●
●
●● ●
●
40 60 80 100 120 140
4060
8010
012
014
0
QQ Plot
theoretical quantiles [mm/day]
empi
rical
qua
ntile
s [m
m/d
ay]
● ●●●●●●●
●●●●
●●●●●●
●●●●●
●●●●●●
●●●●●●●●●●●
●●●
●
●
●
●
●
●
●● ●
●
Left: Time series of extrema.
Middle: Histogram of extrema. Red line: fitted GEV density distribution.
Right: QQ plot. Plot of empirical vs. theoretical quantiles of the observations. The theoretical quan-tiles are estimated with the fitted GEV. Should the dots align on the diagonal (red line), the fit would beperfect. The QQ-plots of 1000 samples randomly drawn from the fitted GEV have a 95% chance of beingwithin the blue lines.
©MeteoSwissKrahbuhlstrasse 58, 8044, Zurichrre150yx; xs 2.3.1; June 10, 2015
Contact: [email protected]
Method information
Documentation ... p. 27
Distribution information
Interpretation ..... p. 15
Data information
Interpretation ...... p. 15
Documentation ... p. 19
Technical Report MeteoSwiss No. 255
10
2 Interpretation
The station-based extreme value analysis is presented on two pages. The first page (shown on page
8 and again in figure 3 for another station, left) states briefly the data employed and the fit quality,
and displays the return level plot, a table of return levels for a selection of return periods, and a table
of return values for the five yearly maxima in the data series. The second page (shown on page 9
for Zurich and again in figure 3 for Geneva, right) contains more detailed information on the data, the
distribution and the estimation.
The extreme value analysis presented on the webpage of MeteoSwiss is based on the generalized
extreme value (GEV) distribution (see section 4), which is used here to describe the behaviour of the
yearly maxima of daily precipitation.
2.1 Return level plots
Figure 1: Return level plotas presented in the extremevalue analysis informationof 1-day precipitation at sta-tion Lugano. The blue linerepresents the return levelestimates based on the esti-mated GEV-distribution; thegreen lines represent their95% confidence interval; andthe y-coordinate of the blackpoints represent data usedfor estimation of the GEV-distribution; the plotting po-sition (x-coordinate) of thepoints is solely determined bythe length of the underlyingrecord (and is not robust).
return period [years]
retu
rn le
vel [
mm
/day
]
1.1 2.33 5 10 20 50 100 200
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
5010
015
020
025
030
0
© MeteoSwiss
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation2 Interpretation
11
Return level plots (figure 1) display the probability that a given value is exceeded. The probability is
expressed in terms of years. Thus a value that has a probability of 1% of being exceeded in any given
year is - in the (very) long term average - expected to be exceeded once in a 100 years. Then, the
value, or “return level”, is said to have a “return period” of 100 years. In fact, return levels are extreme
quantiles: the 100-year return value has a probability of 0.99 of not being exceeded, and therefore
corresponds to the customary 0.99-quantile (see figure 2).
To focus the attention on the behavior of the rarest events, return periods are transformed in such a
way that the relation to the return levels will appear as a straight line for a Gumbel distribution (see
section 4) of the yearly maxima.
0 50 100 150 200
0.0
0.2
0.4
0.6
0.8
1.0
quantile
prob
abili
ty
0.5−quantile0.99−quantile
0.0 0.2 0.4 0.6 0.8 1.0
50
100
150
200
250
probability
quan
tile
0.5−quantile
0.99−quantile
50
100
150
200
250
return period
retu
rn le
vel
0.5−quantile
0.99−quantile
2 5 10 20 50 100 200
Figure 2: Correspondence between cumulative probability distribution function (CDF, left), quantile function (cen-ter) and return level plot (right). The distribution shown here was estimated for Lugano 1-day precipitation ex-tremes (cf. figure 1).
In real life applications, the true probability distribution describing the behavior of extremes is unknown,
and must be estimated from the observations. Due to the finite size of the sample, this estimation is in-
herently uncertain, but the uncertainty can be quantified, and translates to uncertainty of the estimated
return levels (green lines, figure 1).
Remark. It is important to keep in mind that the plotted uncertainties only describe the uncer-
tainty in return level estimates. No statement can be made regarding an uncertainty in return
period based on the uncertainties represented in figure 1.
The return level plot often displays so-called “plotting points” (black points, figure 1). The position of
these plotting points on the y-axis corresponds to the values of the observed sample maxima. Since
the true return period of these events is unknown, however, the position on the x-axis is determined
empirically from the sample size. Thus, for a sample size of 50 years, for instance, the largest obser-
vation will be assigned a return period of approximately 50 years, the second largest a return period
of approximately 25 years, and so on. In other words, it does not represent the local behavior of the
quantity of interest (e.g. daily precipitation) at the particular station under consideration, since the
return period would be the same regardless of the geographical location of the station. Some conclu-
sions can, however, be drawn from the comparison between plotting points and best estimate (blue
line, figure 1): any substantial or systematic disagreement between empirical and GEV estimates, after
allowance for sampling error, suggests an inadequacy of the GEV model.
For a more thorough analysis of the fit, it is preferable to examine the agreement between empirical
and modeled estimates in a quantile-quantile plot, which is easier to interpret correctly (see section
Technical Report MeteoSwiss No. 255
12
4.3.2).
2.2 Reliability
Each extreme value analysis is tested for its reliability, and the verdict is stated on the analysis sheet.
This verdict does not pertain to the inherent uncertainty of the estimation due to the limited sample
size. Rather, it describes the possibility that the assumptions motivating the choice of the statistical
model may not be fulfilled. If the reliability is poor, for instance, the model does not represent well
the observed extreme values. Three degrees of reliability have been defined: poor, questionable, and
good. Their meaning and what to do about them is shown in the following table.
Remark. “The only justification for extrapolating an extreme value model is the asymptotic basis
on which it is derived.” (Coles, 2001, page 3)
“However, if a model is found to perform badly in terms of its representation for extreme values
that have already been observed, there is little hope of it working well in extrapolation.” (Coles,
2001, page 3)
Verdict Interpretation for the user Action
poor Observed extreme values are Do not use statistical model!
badly Use largest observed events or
represented by the model. closest reliable station instead.
questionable Observed values are Careful assessment necessary.
not well Use visual guides (sec. 4.3.1) to decide:
represented by the model. if points line up - proceed as usual;
Underlying assumptions might be violated. if not - treat fit as poor.
good Observed extreme values are
well The statistical model can be used.
represented.
2.3 Example stations
In the following, we illustrate the interpretation of a typical extreme value analysis product for two sta-
tions, Geneve-Cointrin (abbreviation GVE) and Zurich / Fluntern (abbreviation SMA), that are subject
to different climatic conditions in terms of heavy precipitation, and therefore lend themselves well to
highlight concepts and problems in interpreting extreme value analysis.
Geneva is located at the south-western tip of Switzerland, in a flat plain, between the Jura mountains
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation2 Interpretation
13
and the foothills of the Alpine arch, while Zurich lies in the north-eastern part of the Swiss Plateau.
Heavy precipitation is generally due to summer thunderstorms on the background of a weather con-
figuration characterized by moist, south-westerly flow. Geneva is comparatively dry because of its
protected position in the lee of the Jura mountains. Although it is not of direct relevance in the present
context, it is interesting to note that the seasonal cycle of daily precipitation is very modest in Geneva,
with the largest monthly maxima in September; in Zurich on the other hand, it is very pronounced and
peaks in August.
The differences in precipitation regime between Geneve-Cointrin and Zurich/Fluntern are manifest in
the statistical characterization of their extremal behavior. Average heavy precipitation is of similar
magnitude at both stations: we can expect events of 50mm per day to be exceeded on average every
other year, with similar ranges of deviation from this average. Yet, they differ in the behavior of the
very rare events, and thus illustrate two different types of extremal behavior, both typical for one-day
precipitation extrema.
As it turns out, the estimates at both stations are of questionable reliability. As we shall see, however,
the information generated by the statistical model can be used with some confidence in the case of
GVE and not SMA.
Geneve-Cointrin: 420m, 46.25N, 6.13E
Extreme Value Analysis1-day precipitation, 5:40-5:40 UTC
1961 - 2013 (number of missing years: 0)
Block Maxima (GEV). Reliability of results: questionable.
Plot of return levels and their uncer-tainty (ordinate) for a given returnperiod (abscissa).
The best estimate of the return levelsis colored blue. The return level 95%confidence intervals are in green.
The dots represent observations (inthe ordinate) to which empirical return pe-riods (abscissa) have been assigned. Theseempirical return periods depend solely onthe sample size. Here, for instance, thelargest observation is automatically assigneda return period of approx. 53 years.
return period [years]
retu
rn le
vel [
mm
/day
]
1.1 2.33 5 10 20 50 100 200
●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
4060
8010
012
014
0
© MeteoSwiss
Table of the largest annual extrema inthe period analysed.
If two large events occur in the sameyear, only the largest will appear in thistable. The return periods are estimatedwith the fitted GEV distribution.
date precipitation[mm/day]
estimated returnperiod [years]
2002-11-14 92.6 791993-09-09 85.2 411978-08-07 76.1 181975-09-14 75.7 181999-09-25 72.9 14
Table of return levels for a selection ofreturn periods.
The 95% confidence intervals are shown inbrackets.
return return confidenceperiod value interval[years] [mm/day] [mm/day]
2 47.2 ( 43.2 - 51.6 )5 60.5 ( 55.3 - 67.2 )
10 69.1 ( 62.6 - 80.8 )20 77.2 ( 69.0 - 97.2 )30 81.8 ( 72.3 - 108.3 )50 87.5 ( 76.0 - 124.1 )
100 95.2 ( 80.5 - 149.2 )
©MeteoSwissKrahbuhlstrasse 58, 8044, Zurichrre150yx; xs 2.3.1; June 10, 2015
Contact: [email protected]
Distribution Function and estimation methods
• The Generalized Extreme Value distribution (GEV) is fitted to the yearly extrema.
• The distribution parameters are estimated with Maximum Likelihood.
• The confidence intervals are estimated with Profile Likelihood.
Data and data quality
• The raw data is quality-checked, but not homogenized.
• Missing data: None.
Distribution Parameters
• Location: 42.87
• Scale: 11.9
• Shape: -0.02
Additional information
Time series of extrema
time
mm
/day
1960 1970 1980 1990 2000 2010
3040
5060
7080
90
Histogram of Extrema
mm/day
dens
ity
30 40 50 60 70 80 90
0.00
0.02
0.04
0.06
●● ● ●
●●
●●●●●●
●●
●●
●●●
●●●●
●●●
●●●●
●●●
●●●●●●●●
● ● ●
●
●
●
●
●
●●
●
●
30 40 50 60 70 80 90
3040
5060
7080
90
QQ Plot
theoretical quantiles [mm/day]
empi
rical
qua
ntile
s [m
m/d
ay]
●● ● ●
●●
●●●●●●
●●
●●
●●●
●●●●
●●●
●●●●
●●●
●●●●●●●●
● ● ●
●
●
●
●
●
●●
●
●
Left: Time series of extrema.
Middle: Histogram of extrema. Red line: fitted GEV density distribution.
Right: QQ plot. Plot of empirical vs. theoretical quantiles of the observations. The theoretical quan-tiles are estimated with the fitted GEV. Should the dots align on the diagonal (red line), the fit would beperfect. The QQ-plots of 1000 samples randomly drawn from the fitted GEV have a 95% chance of beingwithin the blue lines.
©MeteoSwissKrahbuhlstrasse 58, 8044, Zurichrre150yx; xs 2.3.1; June 10, 2015
Contact: [email protected]
Figure 3: Station-based extreme value analysis information for 1-day precipitation at station Geneve-Cointrin.
Technical Report MeteoSwiss No. 255
14
2.3.1 Geneve-Cointrin
Figure 3 shows the extreme value analysis information of MeteoSwiss for 1-day precipitation recorded
at the station Geneve-Cointrin.
Figure 4: Short summary of extreme value analysis.
Short summary To define the product at hand,
the first first few lines provide succinct information
regarding the data and statistical method (figure 4):
• The station name, here Geneve-Cointrin, its
altitude (420m) and geographical coordinates
(46.25N, 6.13E).
• The MeteoSwiss product name, here extreme value analysis.
• The variable analyzed, here 1-day-precipitation, recorded daily at 5:40 UTC, observed from
1961 to 2013 with no missing years (see section 3).
• The extreme value approach employed: The block maxima approach, where annual maxima
are assumed to follow a generalized extreme value distribution (see section 4).
• A summary of the reliability of the results based on the quality of the fit, here questionable (see
section 4.3.2 and page 37).
Return level plot The return level plot summarizes the extremal behaviour at a glance (figure 5).
A range of 1-day-precipitation amounts (y-axis) is plotted against their exceedance probabilities ex-
pressed in terms of average number of years (x-axis). The extremal behavior can be deduced from
the return level plot (figure 5): The return level estimate (blue line) is a straight line. This means the
probability that a large value will be exceeded decays exponentially. Thus, while any amount of daily
precipitation has a non-zero probability of being exceeded, this probability decreases very rapidly as
the amount increases. Keep in mind that the return level plot has been transformed in such a way that
an exponential decay looks like a straight line.
The strong positive curvature of the upper confidence bound (green line, figure 5) reveals that this
extremal behavior is subject to uncertainty: a slower (polynomial) decay of the exceedance probability
is also quite possible. This means that very severe events would have a higher probability of being
exceeded than in the exponential case. The lower confidence bound is slightly negatively curved - a
bounded distribution is therefore possible but rather unlikely. Which would suggest that beyond a given
daily precipitation amount, the probability that an amount should be exceeded is zero.
The green lines on the return level plot (figure 5) show the 95% confidence intervals. Note that these
confidence intervals only describe the uncertainty of the return levels, expressed in mm.
The dots on the figure are called plotting points. They represent the observed yearly maxima (ordi-
nate) to which have been assigned an empirical return period (abscissa). This empirical return period
depends solely on sample size (here the number of yearly maxima), and is the same for all stations,
regardless of their geographical location.
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation2 Interpretation
15
return period [years]
retu
rn le
vel [
mm
/day
]
1.1 2.33 5 10 20 50 100 200
●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
4060
8010
012
014
0
© MeteoSwiss
Figure 5: Return level plot as presented in the Ex-treme Value Analysis information of 1-day precip-itation in Geneva. The blue line represents the re-turn level estimates based on the estimated GEV-distribution; the green lines represent their 95% con-fidence interval; and the y-coordinate of the blackpoints represent data used for estimation of theGEV-distribution; the plotting position (x-coordinate)of the points is solely determined by the length of theunderlying record (and is not robust).
Tables The table of the largest annual extrema and the table of return levels contain the same infor-
mation as the return level plot. The first (table 1, left) “reads” the return periods corresponding to the
5 largest maxima, the second (table 1, right) “reads” the precipitation amounts corresponding to given
return periods. The relation between the two is dictated by the estimated GEV.
date precipitation[mm/day]
estimated returnperiod [years]
2002-11-14 92.6 791993-09-09 85.2 411978-08-07 76.1 181975-09-14 75.7 181999-09-25 72.9 14
return return confidenceperiod value interval[years] [mm/day] [mm/day]
2 47.2 ( 43.2 - 51.6 )5 60.5 ( 55.3 - 67.2 )
10 69.1 ( 62.6 - 80.8 )20 77.2 ( 69.0 - 97.2 )30 81.8 ( 72.3 - 108.3 )50 87.5 ( 76.0 - 124.1 )
100 95.2 ( 80.5 - 149.2 )
Table 1: Largest 1-day precipitation events with estimated return period (left) and return level estimates to a se-lection of return periods (right) as presented in the extreme value analysis information of 1-day precipitation atGeneve.
Distribution information The estimated GEV distribution is provided via the three estimated param-
eters: the location parameter, here 42.87 mm; the scale parameter, here 11.9 mm; and the shape
parameter, here −0.02. The corresponding probability density function (PDF) is shown in figure 6 as
the red line of the right panel.
Data information The time series of annual maxima (figure 6, left) informs visually about missing
data or possible trends or cycles. Here, there could be a decadal oscillation or a trend towards higher
values at the end of the record. Further inquiries would, however, be necessary to rule out random
effects. The histogram (figure 6, right) shows the frequency with which the annual maxima have been
recorded, and can be understood as the empirical PDF.
Technical Report MeteoSwiss No. 255
16
Time series of extrema
time
mm
/day
1960 1970 1980 1990 2000 2010
3040
5060
7080
90
Histogram of Extrema
mm/day
dens
ity
30 40 50 60 70 80 90
0.00
0.02
0.04
0.06
Figure 6: Time series (left) and histogram (empirical density) (right) of annual maxima with estimated den-sity function (red line) as presented in the extreme value analysis information of 1-day precipitation at stationGeneve-Cointrin.
Estimation information Geneve-Cointrin is an excellent example for the problems experienced when
applying extreme value statistics: the estimated distribution does not fit well, either for the lower tail or
the heavier (but not heaviest, here around 60mm) events, and the fit turns out to be questionable. It is
possible that the lower daily precipitation maxima (perhaps below 30 or 40mm) are not really ”extreme”.
As we shall see in section 4, the maximum should in principle represent the largest of an infinite num-
ber of independent values, and a year may be too short. Thus, fitting the distribution to 2-year maxima,
rather than 1-year maxima, could eventually yield more convincing results. The GEV is also only a
valid approximation if there is no systematic change in process within a year (stationarity assumption).
This assumption is most likely violated by the seasonal cycle in daily precipitation. In such a case,
restricting the analysis to one particular season may improve the results.
Figure 7: Quantile-quantile plot as presented inthe extreme value analysis information of 1-dayprecipitation at station Geneve-Cointrin.
●● ● ●
●●
●●●●●●
●●
●●
●●●
●●●●
●●●
●●●●
●●●
●●●●●●●●
● ● ●
●
●
●
●
●
●●
●
●
30 40 50 60 70 80 90
3040
5060
7080
90
QQ Plot
theoretical quantiles [mm/day]
empi
rical
qua
ntile
s [m
m/d
ay]
●● ● ●
●●
●●●●●●
●●
●●
●●●
●●●●
●●●
●●●●
●●●
●●●●●●●●
● ● ●
●
●
●
●
●
●●
●
●
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation2 Interpretation
17
Remark. What have we learned:
at Geneve-Cointrin, the probability of very extreme precipitation decays exponentially with the
severity of events. Statistically speaking, the data is consistent with a polynomial (therefore
slower) decay of the tail of the GEV distribution. Thus, for each precipitation amount, there exists
a non-negligible probability of exceeding this amount (there exists no upper bound).
The reliability of the fit is questionable. Thus, further investigation would be necessary to find a
statistical model for which the assumptions are less violated. Extrapolation to unobserved levels,
i.e., 100-year and higher return levels, is also not advisable. However, the information can be put
to use as a pragmatic guide providing only very rough orders of magnitude. Given that the fit is
questionable, the uncertainties themselves are not entirely reliable.
The largest 1-day precipitation event (92.6mm) was recorded November 14th, 2002, which corre-
sponds roughly to an 80-year event.
2.3.2 Zurich / Fluntern
The extreme value analysis for 1-day precipitation observed in Zurich / Fluntern is shown on pages 8
and 9. Here, we will highlight differences to the statistical and extremal behavior of the extreme value
analysis for Geneve-Cointrin.
return period [years]
retu
rn le
vel [
mm
/day
]
1.1 2.33 5 10 20 50 100 200
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●5010
015
020
0
© MeteoSwiss
Figure 8: Return level plot as presentedin the Extreme Value Analysis infor-mation of 1-day precipitation in Zurich.The blue line represents the return levelestimates from the estimated GEV-distribution; the green lines representthe 95% confidence interval of the esti-mation; and the y-coordinate of the blackpoints represent data used for estimationof the GEV-distribution; the plotting posi-tion (x-coordinate) of the points is solelydetermined by the length of the underly-ing record (and is not robust).
Technical Report MeteoSwiss No. 255
18
Figure 9: Quantile-quantile plot as presented inthe Extreme Value Analysis information of 1-dayprecipitation at station Zurich / Fluntern.
● ●●●●●●●
●●●●
●●●●●●
●●●●●
●●●●●●
●●●●●●●●●●●
●●●
●
●
●
●
●
●
●● ●
●
40 60 80 100 120 140
4060
8010
012
014
0
QQ Plot
theoretical quantiles [mm/day]
empi
rical
qua
ntile
s [m
m/d
ay]
● ●●●●●●●
●●●●
●●●●●●
●●●●●
●●●●●●
●●●●●●●●●●●
●●●
●
●
●
●
●
●
●● ●
●
Remark. In Zurich, the largest 1-day precipitation event was recorded September 21st, 1968
with 103.1mm. Three events of similar amount (97.8mm, 97.2mm, 97.1mm) have been recorded
08/2007, 08/1978 and 06/1973.
In Zurich, the statistical model does not represent the largest recorded 1-day precipitation events
well, and is therefore unlikely to perform better as regards extremal behavior. The use of any
return level larger than a 5-year event is not advisable. As a first step, a statistical model that takes
into account the annual cycle of heavy precipitation would improve the estimated information.
Meanwhile, as a pragmatic approximation to the extremal behavior, the information about the
largest observed annual maxima can be used.
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation3 Data
19
3 Data
3.1 Stations
For the MeteoSwiss web-page, the observations used were taken at stations of the Swiss National
Basic Climatological Network (NBCN), which groups the most important climatological stations within
the observational network of MeteoSwiss (Begert et al., 2007). Most of these 28 stations have very
long records: data at some of these stations date back to the 19th century. The NBCN stations cover
all of Switzerland (figures 10, 11), and should be to some extent representative of the regional climatic
variations in monthly precipitation.
500
1000
1500
2000
2500
3000
ALT
ANT
BAS
BER
CDF
CHD
CHM
DAV
ELMENG
GRC
GRH
GSB
GVE
LUG
LUZ
MER
NEU
OTL
PAY
RAG
SAE
SBE SIA
SIO
SMA STG
Figure 10: Map of NBCN stations. A list of station names, geographical location, and altitude can be found onpage 42. Black line: line used to define the vertical cross-section across the Alpine ridge, defined as the Rhoneand the beginning of the Rhine valleys, connected to each other and prolonged on either side by straight lines.To the east of the Rhine valley, the straight line is made to pass through the station Scuol.
Technical Report MeteoSwiss No. 255
20
−100 −50 0 50 100 150 200
500
1000
1500
2000
2500
3000
distance to inner−alpine valleys [km]
altit
ude
[m]
ALT
ANT
BAS
BER
CDFCHDCHM
DAV
ELMENG
GRC
GRH
GSB
GVELUG
LUZMER
NEUOTL
PAYRAG
SAE
SBESIA
SIO SMA
STG
Figure 11: Profile of NBCN stations. The vertical cross-section across the Alpine ridge is defined as the distanceto the inner-Alpine valleys (black line in figure 10). The profile of the Alps is computed with the USGS GTOPO30(http://eros.usgs.gov) digital elevation model, and is the minimum air-line distance of each grid-point center to theinner-Alpine valleys. The thick (thin) grey line(s) represent(s) the smoothed median (10% and 90% quantiles) ofthe distances in 100m bins.
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation3 Data
21
3.2 Data quality
The MeteoSwiss observational network encompasses different types of stations. NIME stations ob-
serve precipitation only, and measurements are carried out manually, at 7:30 local time every day.
Climate Stations and SwissMetNet measure a wide range of meteorological parameters. At Climate
Stations, manual precipitation measurements take place daily at 6:00 and 18:00 UTC. At SwissMetNet
stations, precipitation is measured automatically in 10-minute intervals, starting on the hour. Note that
for manual measurements, there is a time window of approximately 70 minutes for the observer to
carry out the reading. The NBCN stations used for the extreme value analyses of precipitation on the
MeteoSwiss WebSite are all Climate Stations.
The measuring instruments of the MeteoSwiss observational network have changed over time (table
2). The first ombrometers measured precipitation by means of an open cylindrical receptacle with a
horizontal cross-section of 400cm2. In the early 20th century, these were replaced with Hellmann rain
gauges with a horizontal cross-section of 200cm2. Both of these systems require manual reading.
At the beginning of the 1980s, about two thirds of the NBCN stations were equipped with automatic
(heated) tipping bucket rain gauges. Since 2012, a further automatic ombrometer has been introduced
that uses a weighing mechanism.
measuring instrument time period reading time
ombrometer 400 cm2 19th / early 20th century 6.00 UTC or 7.30 local time (m)
Hellmann rain gauge 200 cm2 early 20th century till today 6.00 UTC or 7.30 local time (m)
tipping bucket rain gauges since 1980s every 10min., starting on the hour (a)
weighing precipitation gauge starting 2012 every 10min., starting on the hour (a)
Table 2: Precipitation measuring instruments of the MeteoSwiss observational network with manual (m) andautomatic (a) reading.
All of these ombrometers have their weaknesses, and precipitation measurements suffer from both sys-
tematic and non-systematic errors. The Hellmann rain gauges require manual reading at regular times,
and the measurements may have been carried out at incorrect reading times. If the reading time has
been skipped, an accumulated value is provided for the following day, and the amount of precipitation
belonging to the proper time interval is unknown. The tipping bucket systems systematically underes-
timate (overestimate) precipitation at high (low) intensities, while the weighing rain gauges sometimes
yield ghost precipitation due to pressure fluctuations caused by wind, or gradients in temperature.
3.3 Heavy precipitation data
The data analyzed consists of 1-day, 2-day, and 3-day precipitation. The aggregation of 1-day precip-
itation to 2 and 3-day precipitation ignores missing data, i.e. a missing daily sum is treated as zero
precipitation and simply does not contribute to the aggregation. For extreme value analysis with the
block maxima approach (see section 4 for details), only the maximum value per year is considered.
The yearly maximum is computed from the monthly maxima and accepts no missing data, i.e. one
missing month results in a missing year. Monthly maxima of daily precipitation sums are computed if
Technical Report MeteoSwiss No. 255
22
there are no more than two non-consecutive missing days. In other words, two consecutive, or three
non-consecutive, missing days will result in a missing year.
Since long records are essential for the credibility of extreme value analyses, only stations with a mini-
mum of 40 years between 1961 and 2013 were selected (see table 3).
Remark. Requirements regarding observations of 1-day precipitation used for estimating ex-
treme value distribution.
• at least 40 years of observations
• quality checked
• still operated today
• non-homogenized data
• well-documented station history
• at stations with known problems such as SAE, GSB, PIL, WFJ, MLS, COV (abbreviations
are given on page 42) no extreme value distribution is fitted
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation3 Data
23
Station 1-day record 2-day record 3-day record
abbreviation (date) (amount) (date) (amount) (date) (amount)
in [mm] in [mm] in [mm]
ALT 31.07.1874 147.2 31.07.1874 194.7 04.05.2002 204.3
ANT 04.07.1916 185.3 03.05.2002 224.4 16.11.2002 269.5
BAS 25.05.1872 94.8 26.05.1872 139.4 26.05.1872 150.8
BER 14.08.2010 90.3 03.10.1888 139.4 03.10.1888 141.4
CDF 25.08.2002 100.8 19.01.1910 170.4 20.01.1910 206.7
CHD 28.12.1947 96.9 19.01.1910 167.3 20.01.1910 207.6
CHM 25.09.1987 102.9 26.09.1987 175.8 27.09.1987 175.8
DAV 18.11.1874 95.0 19.11.1874 143.0 20.11.1874 198.0
ELM 29.08.1890 147.3 04.07.1891 187.4 15.02.1990 205.6
ENG 31.07.1874 153.8 31.07.1874 226.5 01.08.1874 226.5
GRC 02.11.1968 170.3 15.10.2000 198.4 15.10.2000 226.8
GRH 21.12.1991 151.4 22.12.1991 236.2 22.12.1991 292.8
GSB 12.11.1996 159.6 12.11.1996 273.2 13.11.1996 321.5
GVE 14.11.2002 92.6 09.09.1993 111.3 09.09.1993 129.2
LUG 21.08.1911 262.8 21.08.1911 342.5 22.08.1911 368.4
LUZ 06.06.2002 111.8 06.06.2002 123.3 27.07.1976 133.2
MER 22.08.2005 110.9 22.08.2005 205.0 22.08.2005 218.0
NEU 08.10.1949 116.1 26.09.1987 125.7 20.01.1910 126.1
OTL 26.09.1991 317.9 22.09.1981 377.8 23.09.1981 416.4
PAY 26.09.1987 78.4 26.09.1987 106.8 27.09.1987 106.8
RAG 29.08.1890 175.0 29.08.1890 217.0 31.08.1890 241.0
SAE 22.08.2005 186.7 03.11.1921 290.0 04.11.1921 321.4
SBE 07.08.1978 175.5 15.11.2002 258.5 16.11.2002 367.7
SIA 03.11.2000 108.0 15.11.2002 159.3 16.11.2002 223.0
SIO 14.02.1990 79.3 14.02.1990 138.5 15.02.1990 170.2
SMA 11.06.1876 171.5 12.06.1876 244.5 12.06.1876 272.5
STG 01.09.1881 250.0 02.09.1881 309.1 03.09.1881 338.9
Table 3: NBCN stations with more than 40 available years of daily precipitation data with the records in mm ofthe 1-day, 2-day, and 3-day precipitation.
Technical Report MeteoSwiss No. 255
24
4 Theoretical background: Extreme Value Statis-
tics
Univariate extreme value statistics offers three different approaches to describe extreme events. The
approaches differ in the way they characterize extreme events: via the largest/smallest values; via the
excess of a pre-defined threshold value; or, more generally, via the count of events falling into a pre-
defined set of values. They are called block maxima (BM), peaks over threshold (POT) or point process
(PP) approach, respectively. In the following we will describe the BM approach closely following Coles
(2001).
Here, we first present some general properties of the generalized extreme value distribution (GEV), the
“extremal” distribution that - so extreme value theory tells us - represents the behavior of the maxima.
Then, we explain different methods to infer this distribution from the data at hand, and assess the
goodness of the estimated fit. Finally, we close the section with the notion of return levels, a means to
illustrate the behavior of extreme values.
4.1 The GEV distribution
Using the maxima of a block of observations to characterize the behavior of rare events find its justifica-
tion in extreme value theory. Fisher and Tippett (1928) and Gnedenko (1943) prove that the distribution
of the largest values of any1 random variable converges towards one of three types of limit distributions,
and that these three types are the only possible limit distributions. This law is known as the extremal
types theorem or Fisher-Tippett-Gnedenko theorem. Strictly speaking, the maxima must first be ap-
propriately re-scaled for the limit distribution to be non-degenerate (see Coles, 2001, p.46 for details).
Remark. The extremal types theorem - as does the central limit theorem for sample means - jus-
tifies the practice of approximating the distribution of sample maxima via extreme data only; thus,
it is not necessary to estimate the population’s distribution F in order to describe the behavior of
the population’s rare events.
Remark. Let Mn be a sequence of maxima, an and bn appropriate re-scaling sequences of
constants and G the limit distribution of (Mn − an)/bn, then
Pr{Mn ≤ z} ≈ G∗(z)
1see remark 4.1 and Leadbetter et al. (1983)
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation4 Theoretical background: Extreme Value Statistics
25
with G∗ another member of the GEV family.
In practice, we do not need to estimate the re-scaling sequences, an and bn, but only the distri-
bution G∗.
Remark. For certain distributions F , a non-degenerate limit distribution function does not exist;
this is the case, for instance, for the Poisson and geometric distribution, which both describe
discrete counting processes. For such distributions, the maxima cannot be described by a GEV.
The three types of limit distributions are the Weibull, Gumbel, and Frechet families of distributions.
They can be summarized in one parameterization of the limit model - in exchange for an additional
parameter (von Mises, 1954; Jenkinson, 1955). This generalized limit distribution, known as the GEV
distribution (definition 1), will be used for the subsequent analyses.
Definition 1 (GEV distribution). The family of distributions defined on the set {z : 1+ξ(z−µ)/σ >
0} via
G(z) = exp
(−{
1 + ξ
(z − µσ
)}−1/ξ)with −∞ < µ < ∞, σ > 0, and −∞ < ξ < ∞, is called the generalized extreme value family
of distributions (GEV distribution). The parameter µ is a location parameter, the parameter σ
is a scale parameter and ξ is a shape parameter. The subset of the GEV family with ξ = 0 is
interpreted as the limit as ξ → 0.
Example (GEV distribution). Examples of
Weibull (µ = 45, σ = 11, ξ = −0.1),
Gumbel (µ = 45, σ = 11, ξ = 0), and
Frechet (µ = 45, σ = 11, ξ = 0.25)
probability density functions. These distributions and their samples will be used throughout the
document to illustrate extremal behavior, estimation behavior, uncertainties, and exceedance
probabilities.
They loosely resemble the 1-day-accumulated precipitation annual maxima distributions at Fri-
bourg (not shown), Geneve, and Zurich-Fluntern, respectively.
The cumulative distribution function (CDF) and probability density function (PDF) of a Weibull (red),
a Gumbel (black) and a Frechet (blue) distribution are shown in figure 12. All three have a location
parameter µ = 45, and a scale parameter σ = 11. They differ by their shape parameter ξ, which is
negative (in this example ξ = −0.1) for a Weibull, zero for a Gumbel, and positive ( in this example
ξ = 0.25 ) for a Frechet distribution (see generalized parameterization of definition 1).
Figure 13 shows the PDFs of the three example distributions seen in figure 12, here in separate plots,
with the histograms of the respective samples appearing in figure 14. Figure 14 shows random sam-
ples (of equal length) drawn from these three distributions: these can be imagined as time-series of
annual maxima of daily precipitation.
Technical Report MeteoSwiss No. 255
26
Figure 12: Three extremevalue distribution functions(left: cumulative distributionfunction (CDF); right: proba-bility density function (PDF))belonging to the Weibull (red;ξ = −0.1), Gumbel (black;ξ = 0), and Frechet (blue;ξ = 0.25) distribution fam-ilies, respectively. All threehave the same location pa-rameter (µ = 45) and scaleparameter (σ = 11). 0 50 100 150 200
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035WeibullGumbelFr'echet
Remark. The three different families of distributions are distinguishable by their tail behavior,
and in particular by the rate of decay in the tail:
• The Weibull family, i.e., the subset of GEV distributions with ξ < 0, has a finite upper end-
point (figure 12 red line - also called type III or bounded). Thus, there is 0 probability for
values above that end-point to occur.
• The Gumbel, family, i.e., the subset of GEV distributions with ξ = 0, has an infinite upper
end-point. The tail of the probability density distribution decays exponentially (figure 12,
black line - also called type I or light-tailed). Thus, regardless the value, there is a non-
vanishing probability for a higher value to occur.
• The Frechet family, i.e., the subset of GEV distributions with ξ > 0, has an infinite upper
end-point. The tail of the density distribution decays polynomially (figure 12, blue line - also
called type II or heavy-tailed). Here too, whatever the amplitude of the event, a larger one
has a non-negligible probability to occur.
0 50 100 150 200
0.00
0.01
0.02
0.03
0.04
dgev
(x, l
oc =
45,
sca
le =
11,
sha
pe =
−0.
1)
0 50 100 150 200
0.00
0.01
0.02
0.03
0.04
dgev
(x, l
oc =
45,
sca
le =
11,
sha
pe =
0)
0 50 100 150 200
0.00
0.01
0.02
0.03
0.04
dgev
(x, l
oc =
45,
sca
le =
11,
sha
pe =
0.2
5)
Figure 13: Probability density functions of figure 12 (solid lines) of the Weibull (left, red, ξ = −0.1), Gumbel(center, black, ξ = 0), and Frechet (right, blue, ξ = 0.25) distribution, with a location parameter of 45 and ascale parameter of 11. The histograms (with 50 breaks) show the empirical densities of the three samples oflength 100 seen in figure 14, drawn from the three respective distributions. The vertical dotted line representsthe sample median, while the dashed vertical line at 100 is meant to ease comparison between panels, and withfigure 14.
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation4 Theoretical background: Extreme Value Statistics
27
0 20 40 60 80 100
0
50
100
150
Wei
bull
0 20 40 60 80 100
0
50
100
150
Gum
bel
0 20 40 60 80 100
0
50
100
150
Fre
chet
Figure 14: Three samples of length 100 drawn from the Weibull (top, red, ξ = −0.1), Gumbel (center, black,ξ = 0), and Frechet (bottom, blue, ξ = 0.25) distribution, with a location parameter of 45 and a scale parameterof 11. The horizontal lines indicate, as in figure 13, the sample median (dotted line) and the value 100 (dashedline).
4.2 Inference for the GEV distribution
Inference is - one could say - an educated guess of the distribution, based on the available sample.
Several methods exist to estimate the underlying GEV distribution from the data at hand. Here, we
only present two of them: the Maximum Likelihood (ML) method and the method of L-moments. Both
approaches are used for estimation of the analyses: The default routine is ML estimation. In case no
global maximum is found for the ML estimation, the L-moments estimation is used.
Given the finite size of the sample, the estimates are associated with uncertainties. These describe
a range of values that would also be consistent with the data at hand. For some estimation meth-
ods, theoretical results allow quantification of the estimation uncertainty. Alternatively one can apply
bootstrapping methods to obtain confidence intervals of the estimated parameters or derived quantities.
Remark. As it was customary at the time, the approach of Rothlisberger et al. (1991) was to
first opt for one of the extremal distribution families, and then estimate the scale and location
parameter using regression methods.
In effect, this means selecting a particular paper, with a pre-drawn graph in either log or log-log
scale. Thus, the estimation hinges on the preliminary (educated) guess of the appropriate family.
“But”, as pointed out by (Coles, 2001, page. 47) “there are two weaknesses: first, a technique is
required to choose which of the three families is most appropriate for the data at hand; second,
once such a decision is made, subsequent inferences presume this choice to be correct, and
do not allow for the uncertainty such a selection involves, even though this uncertainty may be
substantial.”
With the adoption of the generalised extreme value distribution parametrization, “the uncertainty
in the inferred value of ξ measures the lack of certainty as to which of the original three types is
most appropriate for a given dataset.“ (Coles, 2001, page. 48)
Notation. Given a sequence of independent random variables X1, . . . , Xn with a common dis-
tribution function F , we denote with Mn = max{X1, . . . , Xn} the maximum of the sequence.
For example, M365 would be the annual maximum of daily observations.
Note that the daily observations should not display seasonal behaviour or dependence on the
Technical Report MeteoSwiss No. 255
28
observations of previous days.
To avoid multiple indices we denote with Zi the maximum Mn of block i, and omit the information
on the block length n. You can imagine {Z1, . . . , Zm} as the set of m annual maxima of daily
precipitation.
4.2.1 Maximum likelihood estimation
Maximum likelihood estimation looks for the (distribution) parameters that are most likely to have led to
the sample at hand, assuming that the largest values of the process of interest are indeed described
by the chosen distribution family, here the GEV distribution.
The likelihood of drawing the recorded values from one particular distribution of the GEV family can
be represented as an n-dimensional surface, where n is the number of parameters of the distribu-
tion. Maximizing this likelihood function or its logarithm (the log-likelihood) yields an estimate of the
parameters.
Parameter estimation The log-likelihood function, `, derives from the distribution function. In our
analyses, the numerical optimization for obtaining the minimum of −` is done with Nelder-Mead at
initial values estimated by L-moments with a fixed ξ = 0.5 (see section 4.2.2 or Hosking (1990) for
details).
Definition 2. Log-likelihood function of the GEV, at z1, . . . , zm,
`(µ, σ, ξ|z1, . . . , zm) =
−m log σ − (1 +1
ξ)
m∑i=1
log
[1 + ξ
(zi − µσ
)]−
m∑i=1
[1 + ξ
(zi − µσ
)]− 1ξ
Notation. To distinguish between theoretical and estimated parameters, one uses the notation
of a small hat above the parameter name, e.g., ξ denotes the theoretical shape parameter, while
ξ denotes the estimated shape parameter.
For each of the samples in figure 14, ML estimation was used to estimate the distribution of the popu-
lation as if the sample were a set of maxima of blocks of observations. Figure 15 shows the histograms
of our samples with the true (solid lines) distribution as in figure 13, as well as the estimated (dashed
lines) distribution of the population. The histogram can be understood as the empirical ”PDF“, i.e.,
without the assumption of a model. The dashed lines show the distribution estimated by maximum-
likelihood from the samples. The vertical lines represent the median of the theoretical (solid line) and
of the estimated (dashed line) distribution. Keep in mind that this is a sandbox example in which we
know the distribution of the population the sample was drawn from (solid lines). In real life, of course,
the population is unknown. We call the estimated distributions (dashed lines) fitted models, because
they are adapted to the observed data.
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation4 Theoretical background: Extreme Value Statistics
29
0 50 100 150 200
0.00
0.01
0.02
0.03
0.04
0 50 100 150 200
0.00
0.01
0.02
0.03
0.04
0 50 100 150 200
0.00
0.01
0.02
0.03
0.04
Figure 15: Histograms of the Weibull (m = 100, left, red), Gumbel (m = 100 center, black), and Frechet sampledata (m = 100, right, blue) seen in figure 14. Overlaid are the PDFs fitted to the sampled data (dashed lines),and the theoretical PDFs (solid lines, see figure 13) of the true distributions from which the data was sampled.
Both the PDFs and the medians of the distributions differ (figure 15), the more so for the heavy-tailed
distribution (right panel). Thus, inferring the underlying distribution from a small sample is difficult, and
the inferred information is inherently uncertain, because it depends on the particular sample we used.
In the setting of maximum likelihood estimation, the uncertainty of the estimates can be reduced by
increasing the number of independent maxima.
For the interpretation of uncertainties it is very important to be precise if one talks about uncertainty
of the estimation, e.g., the uncertainty in the assessment if a coin is fair based on a sample of coin
tosses, or uncertainty about the “prediction” of an event, e.g., the uncertainty if the outcome of tossing
a coin is heads or tails.
The example (figure 15) illustrates the inherent uncertainty of inference with few observed extremes.
Here, we have estimated the distribution from m = 100 years; we can be certain that, by construction,
all assumptions (data quality, representativeness (but sample size), Zi iid, model choice) are fulfilled,
and only the estimation uncertainty still remains (and stochastic uncertainty, which does not play a role
in this section). Of course, this is not the case in real life.
Even though all our assumptions are correct, the fitted model deviates from the true distribution. This
remaining uncertainty is the estimation uncertainty.
Estimation uncertainty ML estimation provides a theoretically founded approximation for the esti-
mation uncertainties. Standard errors of maximum likelihood estimates can be calculated easily. Nev-
ertheless, using the profile likelihood function is recommended, as it allows for asymmetric confidence
intervals (CIs).
Definition 3. Profile log-likelihood confidence intervals for ξ, σ or µ.
For uncertainty estimates of ξ, one fixes ξ = ξ0 for a range of ξ0, e.g. from −1.5 to 2, and for
each ξ0 maximizes the log-likelihood function with respect to µ and σ, given the observations.
Thus one obtains a concave function of ξ with a maximum at ξ. Approximate confidence intervals
can be derived based on the fact that two times the difference in the log likelihood function at ξ0and ξ follows a χ2
1 distribution (see Coles, 2001, section 2.6.6 p.35 for details). The same can
be done for µ and σ.
Technical Report MeteoSwiss No. 255
30
In figure 16 we see the profile log-likelihood functions (black solid lines) for all parameters. The vertical
solid lines represent the true parameters and the three dashed vertical lines indicate the ML estimate,
with the 95% confidence intervals (1.96 × semle) on either side. The profile likelihood CIs are repre-
sented by the dotted vertical lines. They mark the intersection of the profile log-likelihood function with
the critical value of the χ21 distribution (lower blue horizontal line).
40 42 44 46 48 50
−402
−400
−398
−396
Location Parameter
Pro
file
Log−
likel
ihoo
d
6 8 10 12 14 16
−450
−440
−430
−420
−410
−400
Scale Parameter
Pro
file
Log−
likel
ihoo
d
−0.2 0.0 0.2 0.4 0.6
−430
−425
−420
−415
−410
−405
−400
−395
Shape Parameter
Pro
file
Log−
likel
ihoo
d
40 42 44 46 48 50
−414
−412
−410
−408
−406
Location Parameter
Pro
file
Log−
likel
ihoo
d
6 8 10 12 14 16
−450
−440
−430
−420
−410
Scale Parameter
Pro
file
Log−
likel
ihoo
d
−0.2 0.0 0.2 0.4 0.6
−430
−425
−420
−415
−410
−405
Shape Parameter
Pro
file
Log−
likel
ihoo
d
40 42 44 46 48 50
−416
−414
−412
−410
−408
−406
Location Parameter
Pro
file
Log−
likel
ihoo
d
6 8 10 12 14 16
−416
−414
−412
−410
−408
−406
Scale Parameter
Pro
file
Log−
likel
ihoo
d
−0.2 0.0 0.2 0.4 0.6
−480
−460
−440
−420
Shape Parameter
Pro
file
Log−
likel
ihoo
d
Figure 16: Profile log-likelihood functions (solid black curves) for each parameter - location (left), scale (center),shape (right)- estimated from the Weibull (m = 100, red, top), from the Gumbel (m = 100, black, center) or fromthe Frechet (m = 100, blue, bottom) sample data. The blue horizontal solid lines mark the maximum value ofthe profile log-likelihood function (corresponding to the ML estimate) and the critical value (95%) of the χ2
1 dis-tribution. The profile likelihood 95% confidence intervals are indicated by the dotted vertical lines, the maximumlikelihood 95% estimate and confidence intervals are indicated by the dashed vertical lines and the true value isindicated by the solid vertical lines (Weibull in red, Gumbel in black, and Frechet in blue).
As expected, the true values for bounded (Weibull) or light-tailed (Gumbel) type samples are well within
the CIs (regardless of the approximation method) and the two methods for approximating the CIs hardly
make a difference. For the sample of Frechet type, on the other hand, even though the CIs are larger
than in the other two cases, the true value is just barely within the CI based on the standard errors.
Thus, taking confidence intervals into consideration in practical applications is essential if ’capturing’
the true value is of any importance, and extreme value analyses should mention these inherent uncer-
tainties explicitly.
The width of the CI depends on sample size. The example presented here had a sample size of
100. In real life situations, however, we often have only 40 years of observations at our disposal. For
comparison, we have plotted in figure 17 the profile log-likelihood functions for sample size 40 drawn
from the same population distribution as in figure 12. Evidently, the confidence intervals for each
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation4 Theoretical background: Extreme Value Statistics
31
parameter are considerably larger than with 100 years of data. Thus, shorter data records yield larger
CIs.
35 40 45 50 55
−166
−164
−162
−160
−158
−156
−154
Location Parameter
Pro
file
Log−
likel
ihoo
d
6 8 10 12 14 16
−175
−170
−165
−160
−155
Scale ParameterP
rofil
e Lo
g−lik
elih
ood
−0.4 −0.2 0.0 0.2 0.4 0.6 0.8
−168
−166
−164
−162
−160
−158
−156
−154
Shape Parameter
Pro
file
Log−
likel
ihoo
d
35 40 45 50 55
−180
−178
−176
−174
−172
−170
−168
Location Parameter
Pro
file
Log−
likel
ihoo
d
6 8 10 12 14 16
−185
−180
−175
−170
Scale Parameter
Pro
file
Log−
likel
ihoo
d
−0.4 −0.2 0.0 0.2 0.4 0.6 0.8
−185
−180
−175
−170
Shape Parameter
Pro
file
Log−
likel
ihoo
d
35 40 45 50 55
−182
−180
−178
−176
−174
−172
−170
Location Parameter
Pro
file
Log−
likel
ihoo
d
6 8 10 12 14 16
−180
−178
−176
−174
−172
−170
Scale Parameter
Pro
file
Log−
likel
ihoo
d
−0.4 −0.2 0.0 0.2 0.4 0.6 0.8
−210
−200
−190
−180
−170
Shape Parameter
Pro
file
Log−
likel
ihoo
d
Figure 17: Profile log-likelihood functions (solid black curves) for each parameter - location (left), scale (center),shape (right) - estimated from the Weibull sample data (m = 40, top), from the Gumbel sample data (m = 40center), and from the Frechet sample data (m = 40, bottom). The profile likelihood 95% confidence intervalsare indicated by the dotted vertical lines, the maximum likelihood 95% confidence intervals are indicated by thedashed vertical lines and the true value is indicated by the solid vertical lines (Weibull in red, Gumbel in black,and Frechet in blue).
In addition, both the inferred distribution and the approximated confidence intervals depend on the
sample at hand. Figure 18 illustrates the fact that subsamples of equal size (m = 40) result in different
inferred distributions, although they were drawn from the same population. The difference is particularly
striking when drawing from the Frechet family.
Conclusions regarding the temporal evolution of the behavior of extreme events therefore cannot be
based on fitting distributions within different time periods (for instance 1901 - 1950, 1951 - 2000)
and comparing the estimates without taking into account the uncertainties. For such a task, rather
sophisticated tests designed specifically for extremes are necessary.
4.2.2 L-moments estimation
The L-moments (Hosking, 1990) belong to the class of moment estimators, which are quite popular
among hydrologists for the estimation of the GEV parameters. One minor drawback of this method
is the lack of theoretically derived uncertainty estimates, but this can be circumvented by parametric
Technical Report MeteoSwiss No. 255
32
0 50 100 150
0.00
0.01
0.02
0.03
0.04
0 50 100 150
0.00
0.01
0.02
0.03
0.04
0 50 100 150
0.00
0.01
0.02
0.03
0.04
Figure 18: Six fitted PDFs (dashed and dotted lines) estimated from two subsets of Weibull sample data (m =40, left, red), two subsets of the Gumbel sample data (m = 40, 6 center, black), and two subsets of the Frechetsample data (m = 40, right, blue). The theoretical PDFs of figure 13 are represented by solid lines and areconsidered as “truth”.
resampling (see below).
Parameter estimation
Definition 4. L-moment estimation of the GEV parameters (Hosking, 1990)
ξ ≈ −7.8590(
2(3+t3)
− log 2log 3
)− 2.9554
(2
(3+t3)− log 2
log 3
)2with t3 = l3/l1
a,
σ = l2−ξ
(1−2ξ)Γ(1− ξ) with l2 = 1/2 1
(n2)
∑∑i>j(x(i) − x(j)), and
µ = l1 − σξ
(Γ(1− ξ)− 1
)with l1 = 1
n
∑i xi.
al3 = 1/3 1(n3
) ∑∑∑i>j>k(x(i) − 2x(j) + x(k))
Here, in figure 19, we see the L-moments estimates in addition to the maximum likelihood estimates.
The difference between the estimates is very small.
0 50 100 150
0.00
0.01
0.02
0.03
0.04
dgev
(x, l
oc =
lWei
bull$
chi,
scal
e =
lWei
bull$
alph
a, s
hape
= −
lWei
bull$
k)
0 50 100 150
0.00
0.01
0.02
0.03
0.04
dgev
(x, l
oc =
lGum
bel$
chi,
scal
e =
lGum
bel$
alph
a, s
hape
= −
lGum
bel$
k)
0 50 100 150
0.00
0.01
0.02
0.03
0.04
dgev
(x, l
oc =
lFre
chet
$chi
, sca
le =
lFre
chet
$alp
ha, s
hape
= −
lFre
chet
$k)
Figure 19: Three fitted PDFs (dashed lines) estimated from the Weibull sample data (m = 100, left, red), theGumbel sample data (m = 100, center, black), and the Frechet sample data (m = 100, right, blue). The dottedlines are the maximum likelihood estimates as in figure 15. The theoretical PDFs of figure 13 are represented bysolid lines, andare considered as “truth”. The samples of figure 14 appear as histograms with 50 breaks.
Parametric resampling For confidence interval estimations, we choose a parametric resampling
approach with a sample size of 1000.
The idea is to generate surrogate data from the estimated distribution, representing possible, addi-
tional samples. Samples of the same size as the original sample of observations are drawn from the
estimated distribution, which is entirely characterozed by the estimated parameters, hence the name
“parametric resampling”. (Some methods resample the observed data.)
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation4 Theoretical background: Extreme Value Statistics
33
The new, surrogate, samples then serve as starting points for new estimates of the underlying dis-
tribution. From the resulting, here 1000, simulated distribution estimates, we construct 2-sided 95%
confidence intervals, by taking the 2.5% and the 97.5% quantiles of the estimated quantity.
Remark. Estimation uncertainties are not the only uncertainties of an extreme value analysis.
There exist several sources of uncertainty (some of which we cannot quantify):
• the data quality, measurement errors, or the representativeness of the data
• the assumptions that the maxima are independent, identically distributed (i.i.d.)
• the model choice (GEV distribution, sufficient block length)
• the estimation of model parameters
• the interpretation of probabilistic results (stochastic uncertainty / sampling uncertainty)
4.3 Assessing the fit of the model
The validity of the model assumptions usually cannot be checked in a systematic manner. However, it
is possible to compare the empirical distribution G, that makes no assumptions regarding an underlying
model, with the estimated distribution G.
Notation. G empirical distribution, i.e., for each value of the sample, the proportion of the total
number of values in the sample that is below it.
G estimated distribution
This can be done with diagnostic plots or with goodness-of-fit tests. Diagnostic plots are visual guides
to the validity of model assumptions, and will be introduced in the next subsection. Goodness-of-fit
tests quantify the difference between empirical and estimated distribution and are presented in the
subsequent subsection.
4.3.1 Visual guides
Probability-probability (empirical vs. estimated probability) and quantile-quantile (empirical vs. esti-
mated quantile) plots are instruments for visual inspection of fit reliability. The empirical and estimated
probabilities (quantiles) are plotted against each other. Should the estimated distribution describe the
observed data well (up to sampling uncertainty), the points should line up along the line of slope 1 and
intercept 0.
Although they offer essentially the same information, probability-probability (PP) plots and quantile-
quantile (QQ) plots highlight different areas of the distribution.
PP-plots The probability-probability plot plots the estimated probability against the empirical proba-
bility.
Technical Report MeteoSwiss No. 255
34
Definition 5 (probability-probability-plot.). A probability-probability plot consists of the points{(G(z(i)), G(z(i)
): i = 1, . . . ,m
},
where z(i), i = 1, . . . ,m, are the ordered observed values, z(1) ≤ . . . ≤ z(i) ≤ . . . ≤ z(m), G is
the empirical distribution function and G the estimated distribution function.
In the context of extreme values, the PP plot has a weakness: By definition, both the estimated and
the empirical probabilities are bound to approach 1 as the quantile increases. Therefore, the PP plot
offers little (or no) information in the region of most interest.
The QQ plot circumvents this weakness by changing the scale.
QQ-plots The quantile-quantile plot associates to each value of the ordered sample (e.g., a maximum
over a block) the corresponding quantile of the estimated distribution.
By definition, the cumulative distribution function at a given value zp gives the probability p that any
of the values up to (and including) zp may be realized. The value zp is then called the p-quantile:
Pr{Y ≤ zp} = p. The median of a distribution, for instance, is its 12 -quantile or 50%-quantile; if z0.5 is
the median, there is a 50% chance for a realization of the random variable to be below z0.5. If z0.75 is
the 75%-quantile, there is a probability of 75% that any value smaller or equal z0.75 will be be realized.
Thus, if a sample value is such that 75% of sample values are below it, it will be associated on the
QQ-plot with the 75%-quantile of the estimated distribution.
If the model is appropriate for the data at hand, the empirical quantiles of the data should approximately
agree with the estimated quantiles, the more so for a large sample.
For the upper quantiles, empirical quantiles that are systematically smaller than the fitted ones suggest
that the tail of the true distribution may decay more slowly than that of the fitted distribution.
Definition 6 (quantile-quantile-plot.). A quantile-quantile plot consists of the points{(G−1(
i
m+ 1), z(i)
): i = 1, . . . ,m
},
where z(i), i = 1, . . . ,m, are the ordered observed values, z(1) ≤ . . . ≤ z(i) ≤ . . . ≤ z(m), G is
the estimated distribution function.
4.3.2 Goodness-of-fit criteria
We employ goodness-of-fit methods based on empirical distribution statistics, such as the well-known
Kolmogorov-Smirnoff and Cramer-von Mises statistics but also statistics that put more weight on the
tail, i.e., on the rarer events of the distribution.
In a first step, we consider the root-mean-square-error (RMSE) between empirical and estimated dis-
tribution (see table 5). In addition, we assess the appropriateness of the fit of the model via goodness-
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation4 Theoretical background: Extreme Value Statistics
35
●●●●●●●
●●●●●●●●
●●●●●●●
●●●●●●
●●●●●●
●●●●●●●●
●●●●●●●●●●
●●●●●●●●●
●●●●●●
●●●●●●
●●●●●●●●●
●●●●●●
●●●●●●●●●
●●●
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
Probability Plot
Empirical
Mod
el
● ●●●
●●●●●
●●●●●●●●●●
●●●●●●●●●●
●●●●●●●●●●●
●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●
●●●●●●●●●●●●
●●●●●
●●●●●●●●
●●●
● ●
●
30 40 50 60 70 80
3050
7090
Quantile Plot
ModelE
mpi
rical
1e−01 1e+00 1e+01 1e+02 1e+03
4060
80
Return Period
Ret
urn
Leve
l
Return Level Plot
●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●
●●●●●●●●●●●●
●●●●●
●●●●●●●
●
●●●
● ●
●
Density Plot
z
f(z)
20 40 60 80 100
0.00
00.
010
0.02
00.
030
● ●● ● ●●● ●●● ● ● ●● ●●●● ● ●● ● ●● ●● ● ●● ●●● ● ● ●● ●● ●●●●●● ●● ●●●● ● ●● ●● ●● ●●● ●●●● ●●● ●● ● ● ●●● ● ●●● ●● ●● ●● ● ●●● ● ●● ●●● ●● ●● ● ●
Figure 20: Suit of diagnosticplots for the Weibull sample(m = 100) as provided by theR-package ismev. Note thatthe CIs are based on the mlestandard errors and thereforethe upper limit of the CI isunderestimated.
●●●●●●
●●●●●●●●
●●●●●●●●●●●●●
●●●●●●●
●●●●●●●
●●●●●●
●●●●●●●
●●●●●●●●●●●
●●●●●●●
●●●●●●●●
●●●●●●●●
●●●●●●●●
●●●●
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
Probability Plot
Empirical
Mod
el
●●●
●●●●●●●
●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●
●●●●●●●●●●●
●●●●●●●●●
●●●
● ●
●
●
40 60 80 100
4060
8010
012
0
Quantile Plot
Model
Em
piric
al
1e−01 1e+00 1e+01 1e+02 1e+03
4060
8010
0
Return Period
Ret
urn
Leve
l
Return Level Plot
●●●
●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●
●●●●●●●●●●
●●●●●●●●
●●●●
●●
●
●
Density Plot
z
f(z)
20 40 60 80 100 120
0.00
00.
010
0.02
00.
030
● ● ●●● ● ●● ●● ●● ●● ●●● ●●● ●● ●●●● ●● ●● ●● ●●● ●●● ●●●● ●● ●●● ● ●● ●● ●● ●● ●● ● ●● ●● ● ●●● ●●●● ●● ●●●● ●● ●● ●● ●● ● ●● ●●●● ●●● ●● ●●●
Figure 21: Suit of diagnosticplots for the Gumbel sample(m = 100) as provided by theR-package ismev. Note thatthe CIs are based on the mlestandard errors and thereforethe upper limit of the CI isunderestimated.
Technical Report MeteoSwiss No. 255
36
Figure 22: Suit of diagnosticplots for the Frechet sample(m = 100) as provided by theR-package ismev. Note thatthe CIs are based on the mlestandard errors and thereforethe upper limit of the CI isunderestimated.
●●●●●●●●●
●●●●●●
●●●●●●
●●●●●●●●
●●●●●●●●●●
●●●●●●●
●●●●●●●●
●●●●●●●
●●●●●●
●●●●●●●●●
●●●●●●●●
●●●●●●●
●●●●●●●
●●
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
Probability Plot
Empirical
Mod
el
●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●
●●●●●●●
●●●●●●
●● ●
● ●
●
●
40 60 80 100 140
5010
015
0
Quantile Plot
Model
Em
piric
al
1e−01 1e+00 1e+01 1e+02 1e+03
100
200
300
400
Return Period
Ret
urn
Leve
l
Return Level Plot
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●
●●●●●
●
●
Density Plot
z
f(z)
50 100 150 200
0.00
0.01
0.02
0.03
0.04
● ● ●●●●●● ●● ●● ●● ●●●● ●● ●● ●●● ● ●● ● ●●●●●● ● ●● ●● ●●●● ●●●●●●●●● ●●● ● ●●● ● ●●● ● ●●● ●● ●● ●●●●● ●● ●● ●●●● ●● ●●●●● ●● ●●●●●●
of-fit tests. Goodness-of-fit tests test if the overall distance between the empirical distribution and the
estimated distribution is significantly different from zero.
As with the visual guides, one can perform the tests on different scales, and / or put weight on areas
of the distribution one is interested in. As we would like to draw conclusions regarding the behavior of
extremes, we need the right tail of the distribution to be reliable.
To this end, we apply a suit of tests (see table 5 and Luceno (2006)). For each test, the critical values
are obtained by bootstrapping. The null hypothesis (H0) is that the empirical and estimated distribution
are the same.
For the overall decision whether the fitted distribution is suited for extrapolation outside the observed
range, we use the following set of rules:
Verdict Rule
poor any H0 rejected at α = 5%
at least one H0 rejected at α = 20% or
questionable at least 3 H0 rejected at α = 30% or
ξ < −0.15 or
ξ > 0.19
good otherwise
Table 4: Set of rules defining the verdict. Where H0 denotes the null hypothesis of any of the tests given in table5 (i.e., H0: empirical and estimated distribution are the same).
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation4 Theoretical background: Extreme Value Statistics
37
Root-mean-square-error RMSE =√
1/n∑|F (x)− Sn(x)|2 RMSE
Kolmogorov-Smirnoff Dn = supx |F (x)− Sn(x)| KS
Cramer-von Mises W 2n = n
∫∞−∞ |f(x)− Sn(x)|2dF (x) W 2
n
Anderson-Darling A2n = n
∫∞−∞
|F (x)−Sn(x)|2F (x){1−F (x)}dF (x) A2
n
Right-tail Anderson-Darling R2n = n
∫∞−∞
|F (x)−Sn(x)|2{1−F (x)} dF (x) R2
n
Second order right-tail Anderson-Darling r2n = n∫∞−∞
|F (x)−Sn(x)|2{1−F (x)}2 dF (x) r2n
Table 5: Empirical distribution statistics for goodness-of-fit assessment, based on Luceno (2006).
4.4 Return levels
Return levels are the name given to extreme quantiles. A return level to return period m corresponds
to a quantile with probability 1 − 1/m. The equation defining return levels expressed in terms of the
GEV parameters is given in definition 7.
Definition 7 (return level). zp to return period m = 1/p is given by
zp =
µ− σξ [1− {− log(1− p)}−ξ], for ξ 6= 0
µ− σ log{− log(1− p)}, for ξ = 0.
where Pr(Mn ≤ zp) ≈ G(zp) = 1− p, with G a member of the GEV family.
As seen in figures 2 and 12, displaying the probabilities on the original scale of the data does not give
sufficient information on the tails of the distribution. Figure 23 illustrates why the return level plot (see
definition 8) is an intuitive way to present extreme value statistics. It focuses on the tail behavior of the
distribution, by log-log transforming the x-axis. Without any additional tools, one can see if the distri-
bution is bounded or heavy tailed: if the function has a positive curvature (blue line), the distribution
is heavy tailed; if the function has a negative curvature (red line) the function is bounded. In case the
function is linear (black line), the distribution belongs to the Gumbel family.
Definition 8 (return level plot).
{(log(yp), zp) : 0 < p < 1}
with yp = −log(1− p).
Technical Report MeteoSwiss No. 255
38
Figure 23: Theoretical returnlevels (solid lines) of Weibull(red), Gumbel (black), andFrechet (blue) distributions(figure 13). Right panel: sameas left with the estimated re-turn levels (dashed lines -right panel) of the respectivefitted distributions (figure 15).
50
100
150
200
return period
retu
rn le
vel
2 5 10 20 50 100 2002 5 10 20 50 100 2002 5 10 20 50 100 200
50
100
150
200
return period
retu
rn le
vel
2 5 10 20 50 100 200
4.4.1 Return level estimation
The estimation of the return levels is straightforward if the parameters of the distribution are already
estimated: one simply replaces the parameters of equation 7 with their respective estimates and com-
putes the return level of interest.
In general, Pr(Mn ≤ zp) ≈ 1 − p has to be solved for zp. In some cases, for instance when non-
stationarities are modeled explicitly, this must be done numerically as no analytic expression exists.
4.4.2 Uncertainty estimation
As for the parameter uncertainties, the uncertainties of the return level estimates can usually be esti-
mated using the profile likelihood function.
Definition 9. Profile log-likelihood confidence intervals for zp.
For uncertainty estimates of zp, one needs to reparametrize the likelihood function. For instance
one can replace µ with zp + σξ
[1− {− log(1− p)}−ξ
]in the likelihood function. Then one max-
imises the likelihood function with respect to σ and ξ given the observations. Approximate confi-
dence intervals can be derived via the χ2 distribution.
We see in figure 24 that the CIs are not symmetric around the return level estimates: the smaller the
probability, i.e., the longer the return period, the less symmetric the CIs. It is important to keep in
mind that the uncertainties can only be read in one direction: the confidence interval describes the
uncertainty of the return level, and not of the return period.
40
60
80
100
120
return period
retu
rn le
vel
2 5 10 20 50 100 200
40
60
80
100
120
140
160
return period
retu
rn le
vel
2 5 10 20 50 100 200
100
200
300
400
return period
retu
rn le
vel
2 5 10 20 50 100 200
Figure 24: Return level plots of theoretical distribution (solid line), and fitted distribution (figure 15) from samples(figure 14) of size m = 100 (dashed lines) with 95% CI via profile likelihood function (green lines).
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation4 Theoretical background: Extreme Value Statistics
39
Should no analytic expression of the return level exist (as is the case when non-stationarities are
modeled explicitly), uncertainty estimates can be obtained by bootstrapping methods.
4.5 Extreme Value Analysis example
Step by step extreme value analysis of 1-day precipitation recorded in Zurich / Fluntern.
Figure 25 shows the daily precipitation recorded during 40 years at Zurich / Fluntern.
1960 1970 1980 1990 2000 2010
0
20
40
60
80
100
year
mm
/day
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
Figure 25: Daily precipitation time-seriesrecorded at station Zurich / Fluntern be-tween 01.01.1961 and 31.12.2013. Thelargest values per year are circled in red.
For the block maxima approach we need to extract the annual maxima (figure 26).
1960 1970 1980 1990 2000 2010
30
40
50
60
70
80
90
100
year
mm
/day
Figure 26: Time-series of annual maximarecorded at station Zurich / Fluntern between01.01.1961 and 31.12.2013. The climatologicalseason in which the maximum occurs is indi-cated by color: December, January, February(DJF) - black, March, April, May (MAM) - blue,June, July, August (JJA) - red, and (September,October, November) SON - turquoise.
As we can see, most annual maxima occur during the summer season (table 6).
rank amount date
1 103.1 mm 1968-09-21
2 97.8 mm 2007-08-08
3 97.2 mm 1978-08-07
4 97.1 mm 1973-06-23
5 94.2 mm 1999-05-12
rank amount date
6 80.4 mm 1988-08-15
7 75.7 mm 1987-09-25
8 67.2 mm 2000-09-20
9 66.1 mm 1993-07-05
10 60.8 mm 1972-07-21
Table 6: Top 10 annual maximawith the occurrence date, orderedby decreasing amount. The cli-matological season in which themaximum occurs is indicated bycolor: DJF - black, MAM - blue,JJA - red, and SON - turquoise.
To fit the GEV we plug all annual maxima into the negative log-likelihood function and use a numerical
Technical Report MeteoSwiss No. 255
40
optimization routine (in most cases Nelder-Mead) to find the combination of location, scale and shape
parameters that maximize the likelihood surface. The negative log-likelihood maximum of 215.4663 is
reached at location: 44.25, scale: 10.39 and shape: 0.26. These are our maximum likelihood estimates
for the GEV parameters. The corresponding standard errors are 1.65, 1.36, and 0.13, respectively.
We have already discussed the goodness-of-fit of the statistical model in chapter 2, where we showed
the QQ-plot (figure 9).
The corresponding cumulative probability function and probability density function are shown in the left
panel of figure 27a.
0 50 100 150
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150
0.00
0.01
0.02
0.03
0.04
(a) Cumulative probability function (top)and probability density function (bot-tom).
return period [years]
retu
rn le
vel [
mm
/day
]
1.1 2.33 5 10 20 50 100 200
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●5010
015
020
0
© MeteoSwiss
(b) Return level plot as presented in the extreme value analysisinformation of 1-day precipitation at station Zurich / Fluntern with95% confidence interval based on the profile likelihood.
As we are interested in the extremes, we calculate return levels (figure 27b) with their confidence
intervals.
Coming back to the top 10 events, we can now calculate the probability of exceeding these 10 events.
We again express the probability in terms of return periods.
Table 7: Top10 annual max-ima date plusamount and es-timated returnperiod.
date amount return period
1968-09-21 103.1mm 34 years
2007-08-08 97.8 mm 27 years
1978-08-07 97.2 mm 27 years
1973-06-23 97.1 mm 27 years
1999-05-12 94.2 mm 24 years
date amount return period
1988-08-15 80.4 mm 12 years
1987-09-25 75.7 mm 10 years
2000-09-20 67.2 mm 6 years
1993-07-05 66.1 mm 6 years
1972-07-21 60.8 mm 4 years
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentationAbbreviations
41
Abbreviations
CI confidence interval
EVS extreme value statistics
EVT extreme value theory
GEV generalized extreme value
GPD generalized Pareto distribution
` log-likelihood function
mle maximum likelihood estimation
PP probability-probability
QQ quantile-quantile
Technical Report MeteoSwiss No. 255
42
station abbreviation station name longitude latitude altitude [m]
ALT Altdorf 8.62 46.89 438.00
ANT Andermatt 8.58 46.63 1438.00
BAS Basel / Binningen 7.58 47.54 316.17
BER Bern / Zollikofen 7.46 46.99 552.84
CDF La Chaux-de-Fonds 6.79 47.08 1018.00
CHD Chateau-d’Oex 7.14 46.48 1029.00
CHM Chaumont 6.98 47.05 1136.00
DAV Davos 9.84 46.81 1594.16
ELM Elm 9.18 46.92 958.00
ENG Engelberg 8.41 46.82 1035.66
GRC Grachen 7.84 46.20 1605.00
GRH Grimsel Hospiz 8.33 46.57 1980.00
GSB Col du Grand St-Bernard 7.17 45.87 2472.00
GVE Geneve-Cointrin 6.13 46.25 420.00
LUG Lugano 8.96 46.00 273.00
LUZ Luzern 8.30 47.04 454.00
MER Meiringen 8.17 46.73 588.60
NEU Neuchatel 6.95 47.00 485.00
OTL Locarno / Monti 8.79 46.17 366.75
PAY Payerne 6.94 46.81 490.00
RAG Bad Ragaz 9.50 47.02 496.00
SAE Santis 9.34 47.25 2502.00
SBE S. Bernardino 9.18 46.46 1638.70
SIA Segl-Maria 9.76 46.43 1804.00
SIO Sion 7.33 46.22 482.00
SMA Zurich / Fluntern 8.57 47.38 555.95
STG St. Gallen 9.40 47.43 775.66
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentation 43
List of Figures
Figure 1 Return level plot as presented in the extreme value analysis information of 1-day
precipitation at station Lugano. The blue line represents the return level estimates
based on the estimated GEV-distribution; the green lines represent their 95% con-
fidence interval; and the y-coordinate of the black points represent data used for
estimation of the GEV-distribution; the plotting position (x-coordinate) of the points
is solely determined by the length of the underlying record (and is not robust). . . 10
Figure 2 Correspondence between cumulative probability distribution function (CDF, left),
quantile function (center) and return level plot (right). The distribution shown here
was estimated for Lugano 1-day precipitation extremes (cf. figure 1). . . . . . . . . 11
Figure 3 Station-based extreme value analysis information for 1-day precipitation at station
Geneve-Cointrin. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Figure 4 Short summary of extreme value analysis. . . . . . . . . . . . . . . . . . . . . . . 14
Figure 5 Return level plot as presented in the Extreme Value Analysis information of 1-day
precipitation in Geneva. The blue line represents the return level estimates based
on the estimated GEV-distribution; the green lines represent their 95% confidence
interval; and the y-coordinate of the black points represent data used for estimation
of the GEV-distribution; the plotting position (x-coordinate) of the points is solely
determined by the length of the underlying record (and is not robust). . . . . . . . 15
Figure 6 Time series (left) and histogram (empirical density) (right) of annual maxima with
estimated density function (red line) as presented in the extreme value analysis
information of 1-day precipitation at station Geneve-Cointrin. . . . . . . . . . . . . 16
Figure 7 Quantile-quantile plot as presented in the extreme value analysis information of
1-day precipitation at station Geneve-Cointrin. . . . . . . . . . . . . . . . . . . . . 16
Figure 8 Return level plot as presented in the Extreme Value Analysis information of 1-day
precipitation in Zurich. The blue line represents the return level estimates from the
estimated GEV-distribution; the green lines represent the 95% confidence interval
of the estimation; and the y-coordinate of the black points represent data used for
estimation of the GEV-distribution; the plotting position (x-coordinate) of the points
is solely determined by the length of the underlying record (and is not robust). . . 17
Figure 9 Quantile-quantile plot as presented in the Extreme Value Analysis information of
1-day precipitation at station Zurich / Fluntern. . . . . . . . . . . . . . . . . . . . . 18
Figure 10 Map of NBCN stations. A list of station names, geographical location, and altitude
can be found on page 42. Black line: line used to define the vertical cross-section
across the Alpine ridge, defined as the Rhone and the beginning of the Rhine
valleys, connected to each other and prolonged on either side by straight lines. To
the east of the Rhine valley, the straight line is made to pass through the station
Scuol. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Figure 11 Profile of NBCN stations. The vertical cross-section across the Alpine ridge is
defined as the distance to the inner-Alpine valleys (black line in figure 10). The
profile of the Alps is computed with the USGS GTOPO30 (http://eros.usgs.gov)
digital elevation model, and is the minimum air-line distance of each grid-point
center to the inner-Alpine valleys. The thick (thin) grey line(s) represent(s) the
smoothed median (10% and 90% quantiles) of the distances in 100m bins. . . . . 20
Technical Report MeteoSwiss No. 255
44
Figure 12 Three extreme value distribution functions (left: cumulative distribution function
(CDF); right: probability density function (PDF)) belonging to the Weibull (red;
ξ = −0.1), Gumbel (black; ξ = 0), and Frechet (blue; ξ = 0.25) distribution
families, respectively. All three have the same location parameter (µ = 45) and
scale parameter (σ = 11). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Figure 13 Probability density functions of figure 12 (solid lines) of the Weibull (left, red,
ξ = −0.1), Gumbel (center, black, ξ = 0), and Frechet (right, blue, ξ = 0.25)
distribution, with a location parameter of 45 and a scale parameter of 11. The
histograms (with 50 breaks) show the empirical densities of the three samples of
length 100 seen in figure 14, drawn from the three respective distributions. The
vertical dotted line represents the sample median, while the dashed vertical line at
100 is meant to ease comparison between panels, and with figure 14. . . . . . . . 26
Figure 14 Three samples of length 100 drawn from the Weibull (top, red, ξ = −0.1), Gumbel
(center, black, ξ = 0), and Frechet (bottom, blue, ξ = 0.25) distribution, with a
location parameter of 45 and a scale parameter of 11. The horizontal lines indicate,
as in figure 13, the sample median (dotted line) and the value 100 (dashed line). . 27
Figure 15 Histograms of the Weibull (m = 100, left, red), Gumbel (m = 100 center, black),
and Frechet sample data (m = 100, right, blue) seen in figure 14. Overlaid are the
PDFs fitted to the sampled data (dashed lines), and the theoretical PDFs (solid
lines, see figure 13) of the true distributions from which the data was sampled. . . 29
Figure 16 Profile log-likelihood functions (solid black curves) for each parameter - location
(left), scale (center), shape (right)- estimated from the Weibull (m = 100, red, top),
from the Gumbel (m = 100, black, center) or from the Frechet (m = 100, blue,
bottom) sample data. The blue horizontal solid lines mark the maximum value
of the profile log-likelihood function (corresponding to the ML estimate) and the
critical value (95%) of the χ21 distribution. The profile likelihood 95% confidence
intervals are indicated by the dotted vertical lines, the maximum likelihood 95%
estimate and confidence intervals are indicated by the dashed vertical lines and
the true value is indicated by the solid vertical lines (Weibull in red, Gumbel in
black, and Frechet in blue). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Figure 17 Profile log-likelihood functions (solid black curves) for each parameter - location
(left), scale (center), shape (right) - estimated from the Weibull sample data (m =
40, top), from the Gumbel sample data (m = 40 center), and from the Frechet
sample data (m = 40, bottom). The profile likelihood 95% confidence intervals
are indicated by the dotted vertical lines, the maximum likelihood 95% confidence
intervals are indicated by the dashed vertical lines and the true value is indicated
by the solid vertical lines (Weibull in red, Gumbel in black, and Frechet in blue). . 31
Figure 18 Six fitted PDFs (dashed and dotted lines) estimated from two subsets of Weibull
sample data (m = 40, left, red), two subsets of the Gumbel sample data (m = 40,
6 center, black), and two subsets of the Frechet sample data (m = 40, right,
blue). The theoretical PDFs of figure 13 are represented by solid lines and are
considered as “truth”. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentationList of Figures
45
Figure 19 Three fitted PDFs (dashed lines) estimated from the Weibull sample data (m =
100, left, red), the Gumbel sample data (m = 100, center, black), and the Frechet
sample data (m = 100, right, blue). The dotted lines are the maximum likelihood
estimates as in figure 15. The theoretical PDFs of figure 13 are represented by
solid lines, andare considered as “truth”. The samples of figure 14 appear as
histograms with 50 breaks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Figure 20 Suit of diagnostic plots for the Weibull sample (m = 100) as provided by the
R-package ismev. Note that the CIs are based on the mle standard errors and
therefore the upper limit of the CI is underestimated. . . . . . . . . . . . . . . . . 35
Figure 21 Suit of diagnostic plots for the Gumbel sample (m = 100) as provided by the
R-package ismev. Note that the CIs are based on the mle standard errors and
therefore the upper limit of the CI is underestimated. . . . . . . . . . . . . . . . . 35
Figure 22 Suit of diagnostic plots for the Frechet sample (m = 100) as provided by the
R-package ismev. Note that the CIs are based on the mle standard errors and
therefore the upper limit of the CI is underestimated. . . . . . . . . . . . . . . . . 36
Figure 23 Theoretical return levels (solid lines) of Weibull (red), Gumbel (black), and Frechet
(blue) distributions (figure 13). Right panel: same as left with the estimated return
levels (dashed lines - right panel) of the respective fitted distributions (figure 15). . 38
Figure 24 Return level plots of theoretical distribution (solid line), and fitted distribution (figure
15) from samples (figure 14) of size m = 100 (dashed lines) with 95% CI via profile
likelihood function (green lines). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Figure 25 Daily precipitation time-series recorded at station Zurich / Fluntern between 01.01.1961
and 31.12.2013. The largest values per year are circled in red. . . . . . . . . . . . 39
Figure 26 Time-series of annual maxima recorded at station Zurich / Fluntern between 01.01.1961
and 31.12.2013. The climatological season in which the maximum occurs is indi-
cated by color: December, January, February (DJF) - black, March, April, May
(MAM) - blue, June, July, August (JJA) - red, and (September, October, Novem-
ber) SON - turquoise. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Technical Report MeteoSwiss No. 255
46
List of Tables
Table 1 Largest 1-day precipitation events with estimated return period (left) and return
level estimates to a selection of return periods (right) as presented in the extreme
value analysis information of 1-day precipitation at Geneve. . . . . . . . . . . . . . 15
Table 2 Precipitation measuring instruments of the MeteoSwiss observational network with
manual (m) and automatic (a) reading. . . . . . . . . . . . . . . . . . . . . . . . . 21
Table 3 NBCN stations with more than 40 available years of daily precipitation data with
the records in mm of the 1-day, 2-day, and 3-day precipitation. . . . . . . . . . . . 23
Table 4 Set of rules defining the verdict. Where H0 denotes the null hypothesis of any of
the tests given in table 5 (i.e., H0: empirical and estimated distribution are the same). 36
Table 5 Empirical distribution statistics for goodness-of-fit assessment, based on Luceno
(2006). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Table 6 Top 10 annual maxima with the occurrence date, ordered by decreasing amount.
The climatological season in which the maximum occurs is indicated by color: DJF
- black, MAM - blue, JJA - red, and SON - turquoise. . . . . . . . . . . . . . . . . 39
Table 7 Top 10 annual maxima date plus amount and estimated return period. . . . . . . . 40
Technical Report MeteoSwiss No. 255
MeteoSwiss extreme value analyses:User manual and documentationReferences
47
References
Begert, M., G. Seiz, N. Foppa, T. Schlegel, C. Appenzeller, and G. Muller (2007), Die Uberfuhrung
der klimatologischen Referenzstationen der Schweiz in das Swiss National Climatological Network
(Swiss NBCN), Technical Report MeteoSwiss, 215, 43 pp.
Coles, S. (2001), An introduction to statistical modelling of extreme values, Springer Series in Statistics.
Fisher, R., and L. Tippett (1928), On the estimation of the frequency distributions of the largest or
smalles member of a sample, Proceedings of the Cambridge Philosophical Society, 24, 180 – 190.
Gnedenko, B. (1943), Sur la distribution limite du terme maximum d’une serie aleatoire, Annals of
Mathematics, 44, 423 – 453.
Hosking, J. (1990), L-moments: analysis and estimation of distributions using linear combinations of
order statistics, J. R. Statist. Soc. B, 52(1), 105–124.
Jenkinson, A. (1955), The frequency distribution of the annual maximum (or minimum) values of mete-
orological events, Quarterly Journal ofthe Royal Meteorological Society, 81, 158–172.
Leadbetter, M. R., G. Lindgren, and H. Rootzen (1983), Extremes and related properties of random
sequences and processes, Springer Verlag, New York.
Luceno, A. (2006), Fitting the generalized pareto distribution to data using maximum goodness-of-fit
estimators, Computational Statistics and Data Analysis, 51, 904–917.
Rothlisberger, H., J. Zeller, and H. Geiger (1991), Starkniederschlage des schweizerischen alpen- und
alpenrandgebietes.
von Mises, R. (1954), La distribution de la plus grande de n valeurs., vol. II, pp. 271–294, American
Mathematical Society, Providence, RI.
Technical Report MeteoSwiss No. 255
48
Acknowledgment
We thank Thomas Schlegel and Stephan Bader for sharing their knowledge of the Swiss Climate in
numerous discussions.
Technical Report MeteoSwiss No. 255
MeteoSchweiz
Operation Center 1
CH-8058 Zurich-Flughafen
T +41 58 460 91 11
www.meteoschweiz.ch
MeteoSvizzera
Via ai Monti 146
CH-6605 Locarno Monti
T +41 91 756 23 11
www.meteosvizzera.ch
MeteoSuisse
7bis, av. de la Paix
CH-1211 Geneve 2
T +41 22 716 28 28
www.meteosuisse.ch
MeteoSuisse
Chemin de l’Aerologie
CH-1530 Payerne
T +41 26 662 62 11
www.meteosuisse.ch